The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data sources, which creates a vulnerability where malicious instructions embedded in chat logs or social media posts could influence agent behavior.
Ingestion points: The scripts tools/wechat_parser.py, tools/qq_parser.py, and tools/social_parser.py ingest and summarize external files provided by the user.
Boundary markers: The skill attempts to use "Layer 0" persona constraints in prompts/persona_builder.md to define boundaries, but these are instructional rather than technical barriers.
Capability inventory: The skill is authorized to use Bash, Write, and Edit tools, allowing it to execute commands and modify the filesystem based on analyzed content.
Sanitization: The tool outputs are parsed via regular expressions but do not appear to have filters for detecting or stripping prompt injection payloads.
[DYNAMIC_EXECUTION]: The skill generates new executable agent skill definitions (SKILL.md) at runtime within the crushes/ directory. These generated skills are built from data extracted from external logs and user descriptions, which the agent platform then loads and executes.
[DATA_EXPOSURE]: The skill is specifically designed to harvest and aggregate highly sensitive personal information.
Metadata Extraction: tools/photo_analyzer.py uses the Pillow library to extract GPS coordinates, timestamps, and device information from image EXIF data.
Private Log Analysis: The parser scripts extract private conversation patterns, frequency, and emotional content from WeChat and QQ exports.
[COMMAND_EXECUTION]: The SKILL.md utilizes the Bash tool to run local Python utilities in the tools/ directory to facilitate data parsing, version management, and file consolidation.
[EXTERNAL_DOWNLOADS]: The project includes a requirements.txt file that specifies Pillow, a well-known image processing library, as a dependency.

create-crush